{"id":"13466d18-0234-4858-b701-c33e65242e67","arxiv_id":"2507.05797","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DMD modes of stochastic trajectory ensembles, after softmax weighting, provide a normalized spectral fingerprint, a T2* estimate, and stable extrapolation for dephasing dynamics.","lead":"A new data-analysis recipe uses Dynamic Mode Decomposition to turn short ensembles of noisy experimental traces into a frequency fingerprint of the noise plus a decay time, and then uses both to predict the system's future. It is demonstrated on simulated quantum dephasing with white and 1/f noise, but the validation is mostly visual and the code is not released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central spectral-fingerprint claim rests on an unvalidated identification of softmax-transformed DMD mode norms with the noise PSD; the only quantitative check uses a smoothing kernel and a rank selected by the authors, so a non-monotonic PSD test is needed.","rationale":"The reader's verdict of CONDITIONAL is appropriate: the paper is clearly written, the simulations are plausible, and the extrapolation stabilization is an interesting practical contribution. The weakest point, however, is the core spectral-fingerprint identification. Unlike the T2* extraction, which has a direct operational meaning (the real DMD eigenvalue is matched to the observed decay envelope), the softmax(L1) mapping from Section 4.1 has no supporting theory. The validation in Section 4.3 is self-referential: the Gaussian smoothing width is chosen by the authors' formula Eq. (7), and the rank is chosen as the one that gives the best visual agreement. This does not rule out the possibility that the apparent 1/f agreement emerges from the smoothing itself. A non-monotonic PSD test is the minimal check that would settle whether the method is a genuine spectral estimator or just an elaborate curve-matching procedure. I therefore agree with the reader's weakest-assumption identification and see no reason to change the CONDITIONAL verdict on the basis of this pass; the suggested test should be added to the requirements for acceptance. The lack of code/data and error bars are secondary but should also be addressed.","tokens_in":24295,"tokens_out":9263,"duration_ms":110375,"concrete_test":"Run the full DMD pipeline on a synthetic noise ensemble with a known non-monotonic PSD, e.g., two Lorentzian peaks of different heights separated by a spectral notch, or band-limited white noise with a stop band, using the same number of realizations, time window, and rank=20 as in the paper. Compare the softmax weights p_i directly to the true PSD evaluated at the DMD natural frequencies f_i, without the Gaussian smoothing of Eq. (7), using a quantitative metric such as Spearman correlation or Jensen-Shannon divergence after normalization. If the weights do not reproduce the notch or the relative peak heights, the spectral-fingerprint claim is not supported; if they do, the method passes a substantially more discriminating test than Fig. 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that Eq. (5), p_i = softmax(||phi_i||_1), yields a PSD-like spectral fingerprint of the noise—is asserted in Section 4.1 without a derivation or error bound connecting the L1 norms of DMD modes to the spectral density of the driving stochastic process. DMD modes are eigenvectors of a reduced linear propagator fit to observable trajectories; their norms depend on the SVD truncation and the DMD variant used, and the standard reconstruction amplitudes b_i (Eq. 20) are discarded. The only quantitative validation, Section 4.3 and Fig. 5, compares the derived weights with a Gaussian convolution of the true 1/f spectrum whose FWHM is set by the paper's own formula, Eq. (7), and declares rank=20 best by visual agreement. Because the DMD spectrum contains only ten natural frequencies at this rank, the smoothing plus visual match to a smooth 1/f curve is a weak test: a wide family of monotone weight sets would pass. If the softmax(L1) mapping does not track the true PSD, then the spectral-fingerprint result, the 'direct extraction' of T2* in Section 4.4, and the constrained reconstruction using S_i weights in Eq. (8) all lose their physical basis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Dynamical Mode Decomposition (DMD)-based framework for analyzing ensembles of stochastic trajectories. It reinterprets the L1 norms of DMD modes as unnormalized statistical weights, maps them through a softmax function (Eq. 5) to a normalized 'PSD-like spectral fingerprint' of the noise, and claims to extract the coherence time T2* directly from the DMD eigenvalue spectrum (Section 4.4). It then introduces a constrained reconstruction (Eq. 8) that replaces standard DMD amplitudes with these spectral weights and clips eigenvalue magnitudes according to the extracted T2*, with the aim of stabilizing long-time DMD extrapolation. The method is demonstrated on simulated qubit dephasing under 1/f and white noise, including comparisons with Welch and Tikhonov estimators in the Supplementary Material.","tokens_in":24588,"tokens_out":6521,"duration_ms":74490,"significance":"If the central mapping were rigorously established, the framework would be an appealing model-free complement to noise spectroscopy: it operates directly on trajectory ensembles, requires no parametric noise model, and yields a spectral descriptor plus a coherence-time estimate from short records. The paper is clearly written, the benchmark material in the Supplementary Material is a useful contribution, and the Discussion is appropriately cautious about finite-data limitations. However, the central spectral-fingerprint claim currently rests on an unproven identification of softmax-transformed DMD mode norms with the noise PSD, and the only quantitative validation uses a smoothing kernel and a rank criterion chosen after the fact. The extrapolation claim also suffers from a missing amplitude normalization in Eq. (8). These issues are load-bearing and require additional derivation, quantitative validation, and a non-monotonic spectral test before the paper's central claims can be accepted.","major_comments":[{"comment":"The central identification of softmax(||phi_i||_1) with a PSD-like spectral weight distribution is asserted without a derivation or error bound. In addition, the DMD modes in Eq. (19) are eigenvectors of A_r, and the columns of W are only defined up to an arbitrary nonzero scale; unless a normalization convention for W (and hence phi_i) is specified, the L1 norms in Eq. (4) are not gauge-invariant and the softmax weights in Eq. (5) can change by an arbitrary amount under a rescaling of individual modes. Please specify the normalization, derive the claimed connection to the spectral density of the underlying stochastic process, and validate on a non-monotonic spectrum.","section":"Section 4.1, Eq. (5)"},{"comment":"The quantitative validation of the spectral fingerprint is weak. The true 1/f spectrum is convolved with a Gaussian whose FWHM is set by the authors' formula Eq. (7), and rank=20 is declared to give the 'best agreement' by visual inspection. With only ten natural frequencies at this rank and a smooth monotone 1/f curve, a wide family of monotone weight sets would pass the same test, so the comparison does not discriminate the method from alternative monotone weightings. Please provide a quantitative discrepancy metric and test the method on a non-monotonic or band-limited noise spectrum, for which the DMD resolution and the smoothing prescription do not predetermine the outcome.","section":"Section 4.3, Eq. (7), Fig. 5"},{"comment":"The extraction of T2* as the single real eigenvalue of an odd-rank DMD is asserted without justification and without quantitative validation. DMD eigenvalue spectra of finite stochastic data need not contain a single real eigenvalue at odd rank, and even when one is present its identification with the ensemble dephasing envelope is not established by the by-eye comparison in Fig. 6. Please report the extracted T2* against the known simulation value for several ranks and noise realizations, with error bars, and discuss what happens when the real eigenvalue is not present or is not unique.","section":"Section 4.4, Fig. 6 (top)"},{"comment":"The constrained reconstruction in Eq. (8) replaces the DMD amplitudes b_i with the softmax weights S_i, which by construction sum to one. As written, the formula therefore yields X_dmd-cons(t=0) = sum_i S_i = 1, not the initial amplitude of the physical observable, which appears in Fig. 6 only after the true and predicted dynamics are separately normalized to one. This missing overall amplitude factor means the extrapolation reproduces the shape of the decay but not its physical scale. Please include a data-driven amplitude prefactor (e.g., determined from the initial condition or from the reconstruction window) and assess the prediction without per-curve normalization.","section":"Section 4.5, Eq. (8)"}],"minor_comments":[{"comment":"Equations (14) and (15) appear to be duplicated with identical content, including the repeated definition of the double-bracket ensemble average; please remove the duplicate.","section":"Appendix A.1, Eqs. (14) and (15)"},{"comment":"The phrase 'parameter-free choice beta=1' is misleading because the method still relies on user selection of the rank r, the odd-rank choice for T2* extraction, and the smoothing FWHM in Eq. (7) for validation; please soften the wording or specify how these are determined from data-internal diagnostics.","section":"Section 4.1, after Eq. (5)"},{"comment":"The caption states that rank=20 gives the 'closest agreement' but does not report a numerical error metric; adding a quantitative misfit (e.g., relative L2 error or KL divergence) for each rank would make the comparison reproducible.","section":"Section 4.3, Fig. 5 caption"},{"comment":"No data or code availability statement is provided; making the simulation and DMD code available would strengthen the reproducibility of the claims.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the spectral fingerprint is valid: the central claim is under-validated, and the Gaussian-smoothing comparison in Fig. 5 plus the post hoc rank selection do not provide a discriminating test. I would encourage the authors to add a non-monotonic PSD benchmark and a quantitative comparison, and to fix the amplitude normalization in Eq. (8), before resubmission. The manuscript is within scope and the idea is potentially interesting, but acceptance at this stage would be premature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a clearly written, honestly framed paper that proposes a genuinely new way to turn short ensembles of stochastic trajectories into a spectral-weight fingerprint, a coherence time, and a stabilized extrapolation. The softmax-weighted L1 norms of DMD modes (Eq. 5), the odd-rank T2* extraction, and the constrained reconstruction (Eq. 8) are not in the cited literature. The paper deserves a serious referee, but the central spectral claim is not yet demonstrated.\n\nWhat it does well: the reinterpretation of DMD modes as ensemble-level statistical weights is a clever move, and the softmax map does separate dominant modes from a stochastic background in the simulated examples. The authors are careful to call the output a 'PSD-like fingerprint' rather than an absolute spectrum. The supplement benchmark against Welch and Tikhonov inversion is a useful addition, and the extrapolation stabilization looks like a real practical fix for the well-known instability of DMD on noisy data.\n\nWhere it's soft: the mapping from ||phi_i||_1 to actual spectral content of the noise is asserted, not derived. DMD eigenvalues and mode norms depend on rank and SVD truncation, and the standard reconstruction amplitudes b_i are discarded. The only quantitative validation, Fig. 5, convolves the true 1/f spectrum with a Gaussian whose FWHM is set by the authors' own formula (Eq. 7) and then declares rank=20 best by visual agreement. At rank 20 there are only ten distinct natural frequencies, so a broad family of monotone weight distributions would survive that test. The T2* extraction in Fig. 6 is compared by eye to the envelope, with no numerical error metric. There are no error bars, no code or data, and rank sensitivity is discussed qualitatively but not quantified. The stress-test note is right that a non-monotonic PSD, or a case with known spectral features, is needed to see whether the fingerprint actually tracks the spectrum.\n\nBottom line: the paper is a promising contribution for researchers who need model-free noise diagnostics from short trajectories. The central spectral claim is plausible but unproven; the extrapolation part is more convincing. I'd take it to peer review, asking for quantitative validation of the softmax-PSD link, uncertainty quantification, and code/data release. That's the fastest way to know whether this is a real tool or a nice coincidence.","headline":"A clever, clearly written DMD-based noise fingerprint method that deserves peer review, but the central spectral mapping is asserted rather than proven and the validation is too soft to fully back the claims.","tokens_in":25137,"tokens_out":2609,"would_cite":false,"duration_ms":29229,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DMD modes can be read as ensemble-level weights that yield a PSD-like spectral fingerprint, a decoherence time, and stable extrapolation from short noisy trajectory ensembles.","keywords":["dynamical mode decomposition","stochastic trajectory ensembles","noise spectral fingerprint","1/f noise","decoherence time","constrained extrapolation","qubit dephasing"],"falsifier":"Run the pipeline on simulated ensembles with a known non-monotone noise spectrum, such as two well-separated Lorentzian bands of different strengths; if the extracted softmax weights do not place clear peaks at the true band frequencies with roughly the true height ratio, or if the weights visibly track the chosen DMD rank instead of the true spectrum, the spectral-fingerprint claim is refuted.","tokens_in":1788,"feed_emoji":"📊","tokens_out":4069,"duration_ms":84610,"temperature":0.7,"pith_summary":"This paper claims that Dynamical Mode Decomposition, a standard method for extracting oscillatory modes from time series, can be reinterpreted to characterize stochastic noise rather than stationary dynamics. By treating DMD modes as statistical weights over an ensemble of realizations and mapping their $\\ell^1$ norms through a softmax nonlinearity, the method yields a normalized, PSD-like spectral fingerprint that separates white from correlated $1/f$ noise. The same DMD eigenvalue spectrum directly supplies the coherence time $T_2^*$, and a constrained reconstruction that caps eigenvalue magnitudes and uses the learned spectral weights as amplitudes stabilizes DMD extrapolation beyond the measurement window. The demonstration uses simulated qubit dephasing trajectories and requires no parametric noise model, no training, and no specialized control sequences.","feed_headline":"DMD turns noisy trajectories into a spectral noise fingerprint","feed_subtitle":"Model-free pipeline extracts noise spectrum and coherence time from short, noisy records and keeps forecasts stable.","key_machinery":"The load-bearing construction is the reinterpretation of DMD modes as ensemble-level statistical weights. Given matrices $X$ and $X'$ of $n$ stochastic trajectories shifted by one time step, DMD builds a reduced linear operator $A_r$ from the SVD of $X$; its eigenvectors define DMD modes $\\phi_i$, each an $n$-dimensional vector in realization space. Summing each mode's complex magnitudes ($\\|\\phi_i\\|_1$) and passing the result through the softmax function converts the mode norms into positive, normalized spectral weights $S_i$, so that the DMD eigenvalues' frequencies and the weights form a data-driven spectrum. The same eigenvalue set yields a decay time: with an odd rank, one real eigenvalue isolates the ensemble coherence envelope, providing $T_2^*$. For extrapolation, the eigenvalue magnitudes are rescaled to at most $|\\lambda_{T_2^*}|$ while preserving phase, and the amplitudes $b_i$ are replaced by $S_i$, yielding the stabilized reconstruction formula of Eq. (8).","core_discovery":"The central claim is that the standard DMD eigenpairs $\\{\\lambda_i,\\phi_i\\}$ of a stochastic trajectory ensemble carry physical content if $\\phi_i$ is read as a vector in realization space rather than in physical space. Taking the $\\ell^1$-norm of each mode as an unnormalized score, the softmax mapping $S_i = \\exp(\\|\\phi_i\\|_1)/\\sum_j \\exp(\\|\\phi_j\\|_1)$ produces a normalized spectral-weight distribution $\\{\\omega_i, S(\\omega_i)\\}$ that acts as a PSD-like fingerprint of the underlying noise: a sharp peak at the system frequency under weak white noise, a flat background under strong white noise, and low-frequency dominance under $1/f$ noise. The eigenvalue associated with a real DMD mode isolates the coherence envelope, giving $T_2^* = -1/\\mathrm{Re}(\\mu_{T_2^*})$ without fitting a decay function. Finally, the constrained reconstruction of Eq. (8), which caps all eigenvalue magnitudes by $|\\lambda_{T_2^*}| = \\exp(-\\Delta t/T_2^*)$ and weights modes by $S_i$ instead of the initial-condition amplitudes $b_i$, removes the unstable growth that plagues standard DMD extrapolation and tracks the true ensemble-averaged dynamics beyond the analysis window.","pith_inferences":["A testable implication the authors leave implicit: the method should recover known multi-band spectra, not only monotone white and $1/f$ shapes; a two-Lorentzian or bandpass noise test would directly probe whether the softmax weights track spectral shape or merely rank order.","The softmax temperature $\\beta$ offers a tunable contrast knob ($\\beta>1$ sharpens peaks); the authors fix $\\beta=1$ to stay parameter-free, but the ability to vary $\\beta$ makes the fingerprint method robust to the degeneracy they describe, and could be used to quantify uncertainty in the extracted weights across rank choices.","Since the normalization discards absolute power information, the method gives a relative spectral fingerprint; combining it with a separate estimate of total noise power (e.g., from the coherence envelope) would yield an absolute PSD estimate without leaving the DMD framework.","The $T_2^*$ extraction relies on the coherence envelope being associable with a single real DMD eigenvalue; for multi-timescale or non-exponential decay processes, the same idea would produce a distribution of decay rates rather than a single number, which may be a feature rather than a bug."],"forward_implications":["From short, noisy trajectory ensembles, the method yields a normalized spectral fingerprint that separates broadband (white) from correlated $1/f$ noise without any parametric assumption or training.","The decoherence time $T_2^*$ emerges as a distinct real DMD eigenvalue, so a coherence-time estimate can be read directly from the data without fitting a decay model.","The constrained reconstruction of Eq. (8) suppresses the exponential blowup of standard DMD extrapolation, keeping predictions stable and close to the true ensemble-averaged dynamics beyond the measurement window.","Because the analysis is formulated at the trajectory-ensemble level, the same construction applies to any stochastic process with comparable phase-accumulation structure, including classical oscillators with fluctuating instantaneous frequency.","The spectral fingerprint is a finite-data descriptor: its resolution depends on ensemble size, sampling, window length, noise level, and DMD rank, so in practice the extracted features should be checked for stability under rank and window variations."],"supporting_citations":[{"why":"Defines DMD as the decomposition of numerical and experimental data into modes and eigenvalues, the method here reinterpreted.","marker":"[49]"},{"why":"Provides the exact-DMD theoretical basis used to construct the reduced operator $A_r$ and its eigenvalues.","marker":"[52]"},{"why":"Standard DMD textbook formulation and reconstruction formula (Eq. 2) that the constrained scheme modifies.","marker":"[34]"},{"why":"Supplies the $1/f$ noise model and filter-function formalism for qubit dephasing used in the simulations and benchmarks.","marker":"[43]"},{"why":"Introduces the two-state fluctuator model used to generate correlated $1/f$ noise in the simulated ensembles.","marker":"[42]"},{"why":"Describes decoherence in qubits due to low-frequency noise, framing the physical setting the $T_2^*$ extraction addresses.","marker":"[4]"},{"why":"Welch's PSD estimator, the Fourier-based baseline the method is benchmarked against in the supplementary material.","marker":"[54]"}],"fun_headline_variants":["DMD fingerprints noise spectra from trajectory data","Decode noise spectra and decoherence time via DMD","Stable DMD extrapolation with learned noise weights","Extract noise fingerprint and T2* without decay fits","DMD reads noise spectra from short noisy records"],"cache_read_input_tokens":27136,"weakest_assumption_plain":"The entire spectral fingerprint rests on the unproven premise that, after truncating the DMD to a chosen rank, the $\\ell^1$-norms of the DMD modes (before and after the softmax mapping) reflect the relative power of the underlying noise at each frequency; the only quantitative validation is a visual match, at one chosen rank, to the true $1/f$ spectrum after it has been smoothed by a Gaussian whose width is set by the paper's own formula.","fun_headline_variants_meta":{"raw":{"variants":["DMD fingerprints noise spectra from trajectory data","Decode noise spectra and decoherence time via DMD","Stable DMD extrapolation with learned noise weights","Extract noise fingerprint and T2* without decay fits","DMD reads noise spectra from short noisy records"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1546,"prompt_tokens":1050,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":666,"tokens_out":496,"duration_ms":6016,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:19:19.687241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on simulated ensembles with a known non-monotone noise spectrum, such as two well-separated Lorentzian bands of different strengths; if the extracted softmax weights do not place clear peaks at the true band frequencies with roughly the true height ratio, or if the weights visibly track the chosen DMD rank instead of the true spectrum, the spectral-fingerprint claim is refuted.","supporting_citations":[{"cited_title":"Dynamic mode decomposition of numerical and experimental data","cited_arxiv_id":null,"evidence_quote":"Defines DMD as the decomposition of numerical and experimental data into modes and eigenvalues, the method here reinterpreted."},{"cited_title":"Dynamic mode decomposition: data-driven modeling of complex systems","cited_arxiv_id":null,"evidence_quote":"Standard DMD textbook formulation and reconstruction formula (Eq. 2) that the constrained scheme modifies."},{"cited_title":"1/f noise: Implications for solid-state quantum information","cited_arxiv_id":null,"evidence_quote":"Supplies the $1/f$ noise model and filter-function formalism for qubit dephasing used in the simulations and benchmarks."},{"cited_title":"Decoherence and 1/f noise in josephson qubits","cited_arxiv_id":null,"evidence_quote":"Introduces the two-state fluctuator model used to generate correlated $1/f$ noise in the simulated ensembles."},{"cited_title":"Decoherence in qubits due to low-frequency noise","cited_arxiv_id":null,"evidence_quote":"Describes decoherence in qubits due to low-frequency noise, framing the physical setting the $T_2^*$ extraction addresses."}],"review_version":1}